ACCEPTED AT IEEE WIRELESS COMMUNICATIONS MAGAZINE 1

Federated Learning for Industrial Internet of Things

in Future Industries
Dinh C. Nguyen, Ming Ding, Pubudu N. Pathirana, Aruna Seneviratne, Jun Li,
Dusit Niyato, Fellow, IEEE, and H. Vincent Poor, Fellow, IEEE

Abstract—The Industrial Internet of Things (IIoT) offers data volumes. The transfer of a massive volume of IIoT data to
promising opportunities to transform the operation of indus- a remote server for AI training requires much network band-
arXiv:2105.14659v1 [cs.LG] 31 May 2021

trial systems and becomes a key enabler for future industries. width and incurs high communication overhead, both of which
Recently, artificial intelligence (AI) has been widely utilized
for realizing intelligent IIoT applications where AI techniques are unacceptable to time-sensitive IIoT applications such as
require centralized data collection and processing. However, this autonomous driving and real-time healthcare. Importantly, the
is not always feasible in realistic scenarios due to the high reliance on such a central server or third party for data learning
scalability of modern IIoT networks and growing industrial raises critical privacy issues, e.g., user information leakage,
data confidentiality. Federated Learning (FL), as an emerging since these data may contain sensitive information. Moreover,
collaborative AI approach, is particularly attractive for intelli-
gent IIoT networks by coordinating multiple IIoT devices and in the future industries, such a centralized AI architecture
machines to perform AI training at the network edge while may be no longer suitable because IIoT data are not centrally
helping protect user privacy. In this article, we provide a detailed located, but distributed over a large-scale network. Therefore,
overview and discussions of the emerging applications of FL in there is an urgent need to go toward distributed AI approaches
key IIoT services and applications. A case study is also provided for enabling scalable and privacy-promoting intelligent IIoT
to demonstrate the feasibility of FL in IIoT. Finally, we highlight
a range of interesting open research topics that need to be applications at the network edge.
addressed for the full realization of FL-IIoT in industries. Recently, federated learning (FL) [2] has emerged as a
promising solution for realizing cost-effective intelligent IIoT
Index Terms—Federated learning, Industrial Internet of
Things, Industry 4.0, future industries, privacy. applications with improved privacy protection. Conceptually,
FL is a collaborative AI approach that enables training of high-
quality AI models by averaging local updates aggregated from
I. I NTRODUCTION multiple learning clients, e.g., IIoT devices, without the need
Recent advances in communication and smart device tech- for direct access to the local data which thus mitigates privacy
nologies along with the rapid development of industrial in- leakage risks. Moreover, since FL attracts large computation
formatization have promoted the proliferation of the Industrial and dataset resources from a number of IIoT devices to train
Internet of Things (IIoT), with its capability to increase AI models, the IIoT data training quality, e.g., accuracy, would
productivity and efficiency in industries [1]. It is anticipated be significantly improved which might not be achieved by
that IIoT will play an increasingly significant role in the using centralized AI approaches with less data and limited
development of new applications, from smart manufacturing, computational capabilities [3].
smart factory to smart transportation and smart healthcare in Motivated by these appealing characteristics, a flurry of re-
the future industrial revolutions, including Industry 4.0. For search activities combining FL with IIoT in industries has been
example, IIoT can provide innovative solutions to drive smart sparked [2]–[4]. However, these works only focus on certain
manufacturing processes due to its ubiquitous sensing and application domains in IIoT, such as cognitive computing [2],
computation capabilities. industrial artificial intelligence [3], and digital twin-enabled
To realize intelligent IIoT services and applications in in- IIoT [4], while a holistic overview on the use of FL in key
dustries, artificial intelligence (AI) techniques such as machine IIoT services and applications is still missing.
learning (ML) have been widely exploited to train data models. To fill this gap, this article presents and details an integration
Traditionally, AI functions are placed at the cloud or the data of FL and IIoT. Specifically, we present the principle of FL
center for data learning and modeling, which remains some and explain its benefits in IIoT. Then, we provide the state-of-
critical limitations with respect to the rapid increase in IIoT the-art review of the use of FL in important IIoT services, i.e.,
IIoT data offloading and caching, IIoT attack detection, and
Dinh C. Nguyen and Pubudu N. Pathirana are with School of Engineering,
Deakin University, Australia IIoT mobile crowdsensing. Notably, we explore and discuss
Ming Ding is with Data61, CSIRO, Australia the roles of FL in key industrial IIoT applications, including
Aruna Seneviratne is with School of Electrical Engineering and Telecom- smart manufacturing, smart transportation, smart grid, and
munications, University of New South Wales, Australia
Jun Li is with School of Electrical and Optical Engineering, Nanjing smart healthcare. A case study is provided in the area of
University of Science and Technology, China federated healthcare to demonstrate the feasibility of FL in
Dusit Niyato is with School of Computer Science and Engineering, IIoT. We also highlight interesting open issues for future FL-
Nanyang Technological University, Singapore
H. Vincent Poor is with the Department of Electrical Engineering, Princeton IIoT research. The new contributions of this article compared
University, USA. to the state-of-the-art are summarized in Table I.
ACCEPTED AT IEEE WIRELESS COMMUNICATIONS MAGAZINE 2

TABLE I: Comparison with related work and new contributions of this article.
Related works Topic Key contributions
[2] FL for cognitive An FL-based solution for big data driven cognitive computing in Industry 4.0, aiming
computing in Industry 4.0 to improve the performances of poisoning attack resistance, accuracy, and incentive
mechanisms for industrial automation.
[3] FL for industrial artificial An efficient and privacy-enhanced FL solution for industrial artificial intelligence with
intelligence user privacy awareness.
[4] FL and digital twin An architecture of digital twin-enabled IIoT, by using digital twins to capture the
empowered IIoT characteristics of industrial devices to assist FL.
This article FL and IIoT in industries A holistic discussion of the use of FL in IIoT. Particularly,
• We identify and discuss the potential of FL in various key IIoT services, i.e., IIoT
data offloading and caching, IIoT attack detection, and mobile IIoT crowdsensing.
• We then provide a detailed investigation of using FL in four important applica-
tions, namely smart manufacturing, smart transportation, smart grid, and smart
healthcare.
• We then present a case study toward FL-based smart healthcare. Several interesting
open research topics for FL-IIoT in industries are also highlighted.

The remainder of the article is organized as follows. Sec- the server broadcasts the new global update to all clients
tion II describes the key principle of FL and its benefits in IIoT. for optimizing the local models in the next learning round.
We then present the overview of the potential of FL in IIoT 4) Iterated Training: The FL training is iterated until the
services in Section III. In Section IV, we provide an detailed global loss function converges or a desired accuracy is
discussion on the integration of FL into key IIoT applications. achieved. Here, the accuracy of FL can be defined as the
A case study in FL-based smart healthcare is provided in ratio of total accuracies of all clients to the total number
Section V. Finally, Section VI concludes the article. of clients, according to the popular Federated Averaging
(FedAvg) algorithm proposed by Google [3].
II. I NTEGRATION OF FL AND II OT: K EY P RINCIPLE AND
B ENEFITS B. Key Benefits of FL Integration in IIoT
A. Key Principle With its innovative operational concept, FL can offer some
The typical FL-IIoT network is composed of two main important benefits for IIoT applications in industries as fol-
entities: the data clients, e.g., IIoT devices and industrial lows:
sensors, and an aggregator (e.g., an edge server) located at • Data Privacy Enhancement: In the FL system, only the
a base station (BS) or an access point (AP), as illustrated in local updates are required by the central server for the
Fig. 1. FL allows IIoT devices and the server to train a shared AI training, while the local data are kept at local devices,
global model while the raw data are kept at local devices. Here, which thus provides a degree of data privacy. Following
each IIoT user participates in training a shared AI model by the increasingly stringent data privacy protection legis-
using its own dataset and then uploads its local model to the lation such as the General Data Protection Regulation
aggregator for building a new global model. By relying on (GDPR), the capability of protecting user information of
the distributed data training at IIoT devices, the aggregation FL is significant for building sustainable and safe IIoT
server can enrich the training performance without completely systems.
compromising user data privacy [3]. As shown in Fig. 1, the • Low-latency Network Communication: By avoiding the
generic FL-IIoT process includes the following key steps: offloading of huge data volumes to the server, FL can
1) System Initialization and Device Selection: The aggrega- significantly reduce communication costs in intelligent
tion server selects an IIoT task, e.g., road traffic evalu- IIoT networks, e.g., latency, consumed by raw data trans-
ation or healthcare analytic, along with model require- mission. Therefore, FL also helps save much network
ments such as task classification or task prediction, and spectrum resources required for data training.
learning parameters such as learning rate. Moreover, the • Improved Learning Quality: FL attracts large computation
server selects a subset of IIoT devices as the learning and dataset resources from a number of IIoT devices over
clients that should be involved in the FL process. the distributed IIoT network to train AI models. This
2) Distributed Local Training and Updates: Once the subset cooperation would accelerate the convergence rate of the
of the learning clients is determined, the server sends overall training process and improve learning accuracy,
an initial model to the clients to trigger the distributed which might not be achieved by using centralized AI
training. In every communication round, each client trains approaches.
a local model using its own dataset and calculates an Compared to traditional distributed learning [5], [6], which
update. Then, each client uploads its computed update to mostly performs parallel data training without federation, FL
the server for aggregation. can better exploit similar experienced data from distributed
3) Model Aggregation and Download: After receiving all data sources located at distributed IIoT devices, which might
updates from clients, the server aggregates them and otherwise result in ignoring rarely occurring yet important
calculates a new version of global model. Subsequently, exemplars. Hence, FL is able to gain benefits from data feature
ACCEPTED AT IEEE WIRELESS COMMUNICATIONS MAGAZINE 3

IIoT Devices Aggregator

FL Aggregation

Local
Global Model
Server Aggregation Model
Model
Base Access Aggregation
Station Point
Model Model 2
Upload Download
3

Global
Model

Mobile Global
device Model
Mobile Industrial Manufacturing
Devices Machines Robots
Global
Local Local Local Machine Model
computation computation
... computation

Local Data Local Data Local Data

Local Model Local Model Local Model
Robot Iterated
Clients Clients Clients 4
Training

FL-IIoT Architecture Communication in FL-IIoT

Fig. 1: The network architecture and communication process for FL-IIoT.

diversity across the distributed dataset which helps improve resources trains a local model using a noise-added gradient-
the generalizability of the global AI model for better training descent algorithm and collaborates with other entities to build
performance, e.g., enhanced training accuracy. a shared content caching policy.

III. FL FOR II OT S ERVICES B. FL for IIoT Attack Detection

A. FL for the Optimization of IIoT Data Offloading and Industrial devices have become targets of malicious adver-
Caching saries who can attack AI/ML models in smart manufacturing
To meet the ever-increasing computation demands of IIoT and operations, by modifying data inputs or changing learning
users and industrial operators in Industry 4.0, data offloading network weights which can lead to erroneous predicted out-
has been widely regarded as an efficient solution which enables puts. Many solutions have been proposed to cope with attacks
IIoT devices and machines to offload their data tasks to on IIoT devices such as ensemble diversity or adversarial
resourceful edge servers. In this context, FL can be used training, but they are mostly applied to a specific type of
to implement offloading optimization where multiple IIoT attack and do not scale well to distributed IIoT networks. FL
devices like actuators in smart manufacturing work as intelli- has emerged as a strong alternative to provide collaborative
gent agents to collaboratively train an AI model to learn the intelligence for IIoT systems with the ability to detect and
policy of offloading industrial data, e.g., production-related prevent various attacks for safe industrial processes. Enabled
data packages. This solution not only enhances data privacy by the privacy-promoting feature of FL, a federated attack
due to the distribution of data learning in different IIoT detection and defense solution is built in [9] where each IIoT
devices but also mitigates the computation burden posed on the machine joins to run a deep neural network locally, in order
industrial system in the centralized offloading architecture. For to retrain the threat model to fight against adversaries. An
example, FL can support data offloading for the transportation example for federated attack detection in IIoT is illustrated
industry [7], where each vehicle collaboratively performs data in Fig. 2. Here, each IIoT device first produces adversarial
training for offloading optimization. It prevents sharing actual samples to create a retraining set that is then used to build
data and thus helps address privacy concerns of vehicle drivers. a local attack detection model. Subsequently, the trained
Data offloaded from IIoT devices can be cached by edge gradient is transmitted to the cloud server for aggregation and
servers where FL can play an important role in establishing synchronization to produce a shared model, and this process
intelligent caching policies, in order to cope with the ex- is iterated in several communication rounds until the attack
plosive growth of industrial data in modern IIoT networks. model converges. In this regard, the built model can effectively
As shown in [8], FL is very useful to build proactive data detect attacks thereby building a strong defense solution in
caching schemes in urban informatics where an IIoT-based the IIoT network. Through simulations with MINIST datasets,
transportation system is created by the federation of vehicular the FL-based approach can achieve a high attack detection
entities, including macrobase stations, road side units, and accuracy of 87.8%, compared to 73.8% in the standalone
moving vehicles. Here, each vehicle equipped with caching method.
ACCEPTED AT IEEE WIRELESS COMMUNICATIONS MAGAZINE 4

Cloud Server
Aggregation
Attack Monitor Device 1 Gradient

...
Update

...

...
...
Device 2 Gradient

.
.
.
.
.
.

.
.
.
.
.
.

...

...
Attack Attack Attack

.
.
.

.
.
.
scheme scheme scheme Device K Gradient

.
.
.

...
Old Model

Attack Library Federated Defense New Model

Gradient Update
Attack Scheme Assignment

Natural Natural Natural

Samples Samples Samples
..

..

..
.

.
Retrain Retrain Retrain
..

..

..

..

..

..
.

.
...
..

..

..
.

.
..

..

..

..

..

..
.

.
..

..

..
.

.
Adversarial Adversarial Adversarial
Samples Model Samples Model Samples Model
IIoT Device 1 IIoT Device 2 IIoT Device K

Fig. 2: Federated attack detection and defense in FL-based IIoT networks.

C. FL for IIoT Mobile Crowdsensing data sharing among manufacturers and factories which is not
With the rapid development of IIoT, mobile crowdsensing is an ideal solution due to growing user privacy concerns. FL
designed to take advantage of pervasive industrial devices for can realize intelligence for industrial systems without data
sensing and collecting data from physical environments. For exchange, by the collaborative data learning of distributed
example, operators in smart factories can make decisions based industrial devices and machines. Given the fact that there
on environmental information collected from ambient sensors are diverse industrial services in industries, e.g., production
distributed across the whole factory, e.g., sensing abnormal monitoring with robots, product assembly with automatic ma-
machinery noise to monitor operating status of machines. To nipulator arms, package delivery and logistics with vehicles,
realize intelligent mobile crowdsensing, centralized AI/ML it is desired to develop a multiple FL services solution to
techniques are used which usually require direct access to deal with different industrial services in the co-working IIoT
user data, which in turn makes the data vulnerable to privacy ecosystem, as illustrated in Fig. 3. Each group of industrial ma-
leakage. Moreover, the use of a central server to handle all chines participates in the collaborative AI training using their
sensing industrial data is not a scalable solution, making local industrial datasets within their working environments,
it hard to cope with massive data volumes in large-scale e.g., machinery fault data in production lines and productivity
industrial systems. FL is a promising tool to accelerate the information in the assembly process, before offloading the
learning and training for crowdsensing models. As an example, learned parameter to the cloud via its edge server. Here, each
the study in [10] shows an FL-based mobile crowdsensing virtual machine in the cloud will compute a global model of
scheme, with a focus on privacy-enhancing extreme gradient the industrial service that it manages. In this regard, a multiple
boosting with the cooperation of multiple clients like industrial FL services solution is realized, and all industrial machines
machines. A secure gradient aggregation algorithm is designed in different service lines can benefit from the exchanged
by integrating homomorphic encryption with secret sharing, knowledge.
which prevents the central server from guessing decryption
result before operating aggregation. Simulations reveal a high
B. FL for Smart Transportation
accuracy rate of 98%, and a reduction of 23.9% runtime, and
33.3% communication latency for gradient aggregation. Recent advances in sensing and communication technolo-
gies along with the growth of data volume from road cameras,
IV. FL FOR II OT A PPLICATIONS embedded devices, and vehicular sensors have empowered
vehicular networks. AI/ML has been adopted to realize intel-
In this section, we present the use of FL in IIoT applications
ligent transportation systems (ITS) where massive vehicular
in details.
data are often processed at a data center before sending back
to vehicles and roadside units. However, this approach remains
A. FL for Smart Manufacturing some critical issues such as privacy leakage and communica-
Smart manufacturing refers to the integration of intelligence tion overhead caused by raw data sharing. FL can support
into manufacturing processes where AI techniques play im- ITS by running ML models directly at vehicles based on their
portant roles in learning big data generated from industrial datasets such as road geometry, collision avoidance, and traffic
machines for process modeling, monitoring, prediction and flow [11]. A cloud server can be employed to aggregate the
control in production stages. The AI functions often require local updates of all vehicles to make overall decisions on the
ACCEPTED AT IEEE WIRELESS COMMUNICATIONS MAGAZINE 5

A multiple federated learning services

VM: Virtual Machine VM VM VM model for smart manufacturing in the
co-working IIoT ecosystem with
Cloud different industrial services, e.g.,
production monitoring, product
Model Model
server assembly, package delivery and
download upload logistics.
Global Model M1 Global Model M2 Global Model M3

Model
Aggregation

Edge Edge Edge

Server Server Server
Training Training
Training
Local Local Local
Production Model M1 Product Package Delivery
Model M2 Model M3
Monitoring Assembly and Logistics

Industrial Manipulator Delivery

Robots Arms Vehicles

Fig. 3: Federated learning for smart manufacturing.

traffic flow. The use of massive data from multiple vehicles In this way, user information such as energy preference and
and huge computation capability of all participants helps home addresses is not revealed to the server which promotes
provide better traffic prediction outcomes, which cannot be privacy protection. Simulations are conducted with over 800
met by using centralized ML techniques with less dataset and homes in the United States, showing a reduction in networking
limited computation. FL can also support privacy-enhanced load, compared to standalone learning approaches. Further,
smart transport logistics, e.g., package delivery services. In due to the increasing privacy concerns, the shareholders of
this context, postal operators and customers can federate to distributed electric generators and consumers may not willing
run a shared ML model for delivery time prediction based to provide information of their electricity loads/consumption,
on their local data sources, e.g., traffic conditions, drivers’ but these datasets would be critical to the safe operation of
behaviors, and weather, for delivery latency optimization and smart grids. FL can help address this issue by allowing the
thus facilitating logistic activities. Moreover, FL can be also distributed participants to collaboratively learn the patterns of
used in Unmanned Aerial Vehicles (UAVs)-based vehicular electricity generation/consumption, without sharing raw data
networks where UAVs can be employed as mobile FL clients to each other. FL is also an efficient solution to bring together
to join the collaborative model training via aerial links with different stakeholders from energy systems (heat, cool, gas,
an ITS entity such as a road side unit. This federated mobile etc.), aiming to achieve privacy-enhanced energy information
model can enable interesting ITS services such as dynamic exchange in the electricity production ecosystem.
traffic prediction and road weather monitoring, in which
ground-based communications are unavailable. D. FL for Smart Healthcare
In the past few years, AI/ML technologies have been widely
C. FL for Smart Grid used in the healthcare sector to gain insights into health issues
Smart grid plays an integral part in building smart city ar- and diseases by learning digital medical information extracted
chitectures which not only provides energy resources to smart from electronic health records (EHRs) for facilitating diag-
city applications such as transportation, manufacturing, but nosis and severity assessment as well as promoting medical
also has impacts on environmental, security, and social aspects research. One of the challenges in such traditional AI tech-
in Industry 4.0. FL can enable intelligent solutions for smart niques is privacy leakage during data analytics. Indeed, com-
grid management and energy transmissions in a decentralized pared to other domains, data in healthcare systems are highly
manner while helping promote privacy. FL is used to establish sensitive subject to health regulations. Moreover, collecting
federated predictive power schemes in a network of edge data a large volume of clinical datasets from isolated medical
centers for smart grid [12]. In this case, each edge equipment centers is a critical challenge. FL can provide much more
such as a smart meter cooperatively trains AI models, e.g., efficient solutions for data learning and potentially reshapes the
neural networks, using its own electrical consumption data current intelligent healthcare systems by providing intelligent
while the edge server coordinates local updates to build a healthcare services while promoting well user privacy based on
global model to estimate future household electrical demands. the cooperation of multiple entities such as health users and
ACCEPTED AT IEEE WIRELESS COMMUNICATIONS MAGAZINE 6

healthcare providers across medical institutions. Indeed, FL generator in each global round and exchange them with
can offer flexible and privacy-promoting EHRs management the cloud server for aggregation. For every global epoch,
solutions [13], by facilitating the cooperation of multiple each institution collaboratively trains its discriminator and
hospital institutions to perform health data analytics without generator. Specifically, the generator produces minibatchs of
the need for EHRs data sharing. Moreover, FL with its privacy- fake samples from the noise probability distribution. Also,
enhanced nature can promote secure healthcare cooperation for the discriminator samples minibatchs of real data from the
better medical service delivery, by allowing for aggregating actual image distribution. Then, each institution updates si-
the model updates from separate hospital organizations with multaneously the discriminator and generator by ascending
multiple devices, e.g., magnetic resonance imaging (MRI) its stochastic gradients to update its own weights. After local
scanners, to build stronger AI models for medical tasks, such training, all institutions transmit the learned updates to the
as medical imaging. cloud server for model averaging, while actual COVID-19
images are kept at local institutions which thus ensures data
V. C ASE S TUDY privacy. Then, the cloud server broadcasts the new global
We present a case study on FL-IIoT, by designing an FL- updates to all institutions for the next round of GAN learning.
health system for COVID-19 detection [14]. In the pandemic, The FL process is iterated until the global loss function
collecting sufficient data for training becomes challenging with converges with a desired accuracy.
privacy concerns caused by public data sharing. Hence, we
propose a new FL scheme to generate realistic COVID-19
images for facilitating privacy-enhanced COVID-19 detection C. Illustrative Results
with generative adversarial networks (GANs) [15]. Compared We report simulation results obtained when training a
to the traditional FL scheme [14], our advanced FL solution COVID-19 dataset [14] of total 620 X-ray images in three
can achieve federated data augmentation for generating high- classes: COVID-19, normal, and pneumonia in an FL system
quality synthetic COVID-19 images that can enhance the with five institutions. By using the proposed FL model, we
training performances with privacy awareness. The details of generate 1500 synthetic X-ray images which are then com-
our FL design will be provided in the following. bined with an actual dataset for COVID-19 classification. We
use a CNN-based classifier with three convolutional layers
A. System Model and Adam optimizer, and the configurations of GANs are
We consider a system model for FL-based COVID-19 shown in Fig. 4. We evaluate our approach and compare
detection as illustrated in Fig. 4, including a set of medical with state-of-the-art schemes, including the standalone scheme
institutions and a cloud server. Each institution participates in (training dataset at only an institution without federation), the
the FL process using its own COVID-19 image dataset, e.g., standalone scheme with GAN [15], the FL scheme without
X-ray images, to build a global GAN with the cloud, aiming GAN [14], and the centralized scheme.
to generate high-quality synthetic COVID-19 images for im- In Fig. 5(a), we compare the discriminator loss of our
proving the overall COVID-19 detection. Specifically, at each advanced FL scheme and the standalone scheme with GAN
institution we design a GAN consisting of two components, [15]. It can be seen that the the performance of the standalone
namely a generator and a discriminator based on CNNs which scheme cannot achieve its optimum due to the lack of access to
alternatively train via a min-max game [15]. Given a noise the full dataset. Meanwhile, our advanced FL scheme can learn
sample from a standard Gaussian distribution, the CNN-based over the entire data span from distributed datasets which is able
generator learns to generate a fake COVID-19 image data to extract better image features for efficient data augmentation.
point. Moreover, we design another CNN as a discriminator We then investigate the detection accuracy for different FL
at each institution which tries to classify the real COVID-19 schemes and our advanced FL scheme, where the standalone
image data point against the one produced from the generator. scheme is used as the baseline. As shown in Fig. 5(b), the more
The discriminator outputs 1 if the input is real data samples or participating institutions in data training, the higher accuracy
0 if the input is fake data samples. Accordingly, the generator achieved. The intuition behind this observation is the improved
and the discriminator at each institution interact to obtain the image feature learning efficiency thanks to the use of diverse
optimal parameters in a fashion that the generator can generate data sources. Nevertheless, the accurate rate of our FL scheme
the fake COVID-19 image data distribution close to the real is the best among all approaches and is close to the centralized
image data as much as possible to fool the discriminator scheme.
while the discriminator tries to differentiate between fake and Additionally, we compare the accuracy performance of our
real image samples. As a result, the generator can synthesize scheme with other COVID-19 detection schemes, as indi-
realistic COVID-19 image samples which are similar to the cated in Fig. 5(c). Our advanced FL scheme can signifi-
real COVID-19 image data after a training period, aiming to cantly improve the accuracy performance due to its GAN
achieve an efficient data augmentation for later classification and federated learning combination. Our scheme yields the
tasks. highest accuracy of 0.963 after 200 iterative epochs, while
other schemes including the FL scheme without GAN, the
B. FL Training for COVID-19 Detection standalone scheme with GAN, and the standalone scheme
Each institution joins the FL training with the cloud server, without GAN have lower performances, with 0.922, 0.856,
by updating the parameters of the discriminator and the and 0.705, respectively.
ACCEPTED AT IEEE WIRELESS COMMUNICATIONS MAGAZINE 7

Model Model Cloud Server Training at an institution

Upload Download Model Generator
Convtr Convtr
Aggregation G Convtr Convtr Convtr
D: Discriminator (Convtranspose)
G: Generator
D Noise
: Discriminator parameter
: Generator parameter
;
;

Institution 1 Institution 2 Institution N 64x1 4x4x256 4x8x256 8x8x256 16x8x128 16x16x128 64x64x1

Conv Conv Conv Conv Conv

G G G Flatten
... D Training
D Training D Training

Real
4x128x 4x256x 4x256x 4x128x 4x128x 64x64 image
256 256 256 256 128 x1
COVID
X-ray
Normalize Softmax
Images

Fake/Real COVID Normal Pneumonia

Local Data Local Data Local Data Discriminator

Fig. 4: Our advanced FL model for COVID-19 detection.

8 1 1
Our advanced FL scheme
7
Discriminator loss performance

Standalone GAN scheme 0.9

0.8
6
0.8
0.6
5
Accuracy

Accuracy
0.7
4 0.4
0.6
3 Standalone scheme
0.2 FL scheme with 2 institutions Standalone scheme without GAN
FL scheme with 3 institutions 0.5 Standalone scheme with GAN
2
FL scheme with 4 institutions FL scheme without GAN
0 Centralized scheme 0.4 Centralized scheme
1
Our advanced FL scheme with 5 institutions Our advanced FL scheme

0 0.3
0 100 200 300 400 500 50 100 150 200 50 100 150 200
Training epochs Epochs Epochs

(a) Performance of discriminator loss. (b) Performance of COVID-19 detection accu- (c) Performance of COVID-19 detection accu-
racy. racy.
Fig. 5: Performance comparison of different approaches for COVID-19 detection.

VI. C ONCLUSIONS AND O PEN R ESEARCH T OPICS their computational resources for training. Unfortunately,
this requirement is not always met due to the resource
This paper provided a detailed overview on the integration
constraints of certain IIoT devices with weak computation
of FL into IIoT in industries. The roles of FL in important
capacities, e.g., industrial wearable sensors, which can
IIoT services and applications were identified and analyzed.
cause significant delays in model aggregation. Thus,
The feasibility of FL in IIoT was demonstrated via a case study
resource-aware FL algorithms and resource allocation
and simulations. Several interesting open research topics for
solutions should be considered in FL-IIoT system design.
FL-IIoT in industries are highlighted as follows:
• Communication Issues in FL-IIoT: Communications in • Economic Issues in FL-IIoT: In FL-IIoT, when an indus-
FL-IIoT training in both uplinks and downlinks rely trial user serves as training nodes, how to encourage them
heavily on the level of interconnection among machines, to join the FL process is a key challenge. A user may
AI software, and the computation server. This communi- not be willing to devote its resources to perform data
cation network also differs from traditional ones due to training if it does not have much economic benefits to
environmental constraints, such as high temperature and compensate the consumption of computational resources.
corrosive substances in manufacturing processes. Further, Incentive mechanisms such as credit-based support and
the high frequency bands, e.g., above 2.4GHz for WiFi revenue payment are highly needed to attract more users
networks, which are essential for low-latency FL com- to join FL training which also enhances the robustness of
munications, may be not available in realistic industrial industrial FL-IIoT systems.
environments like hospitals. New designs of efficient
communication protocols specific to IIoT settings are ACKNOWLEDGMENTS
desired to facilitate the FL training. This work was supported in part by the CSIRO Data61,
• Resource Management Issues in FL-IIoT: The concept of Australia, and in part by U.S. National Science Foundation
FL-IIoT mostly relies on scalable data parallelism and on- under Grant CCF-1908308. The work of Jun Li was supported
device training at IIoT devices. To achieve a synchronous by National Natural Science Foundation of China under Grant
update at the server, all IIoT devices need to devote 61872184.
ACCEPTED AT IEEE WIRELESS COMMUNICATIONS MAGAZINE 8

R EFERENCES Dinh C. Nguyen is currently pursuing the Ph.D. degree at the School
of Engineering, Deakin University, Victoria, Australia. He is also affiliated
[1] E. Sisinni, A. Saifullah, S. Han, U. Jennehag, and M. Gidlund, “In-
with the Information Security and Privacy Research Group, CSIRO Data61,
dustrial Internet of Things: Challenges, Opportunities, and Directions,”
Docklands, Melbourne, Australia. His research interests focus on federated
IEEE Transactions on Industrial Informatics, vol. 14, no. 11, pp. 4724–
learning, blockchain, Internet of Things, and edge computing.
4734, Nov. 2018.
[2] Y. Qu, S. R. Pokhrel, S. Garg, L. Gao, and Y. Xiang, “A Blockchained
Federated Learning Framework for Cognitive Computing in Industry 4.0 Ming Ding is currently a Senior Research Scientist with the CSIRO Data61,
Networks,” IEEE Transactions on Industrial Informatics, pp. 1–1, Jul. Sydney, NSW, Australia. His research interests include information technol-
2020. ogy, data privacy and security, machine learning and AI. He has authored
[3] M. Hao, H. Li, X. Luo, G. Xu, H. Yang, and S. Liu, “Efficient and over 100 articles in IEEE journals and conferences. He is an Editor of
Privacy-Enhanced Federated Learning for Industrial Artificial Intelli- the IEEE Transactions on Wireless Communications and the IEEE Wireless
gence,” IEEE Transactions on Industrial Informatics, vol. 16, no. 10, Communications Letters.
pp. 6532–6542, Oct. 2020.
[4] W. Sun, S. Lei, L. Wang, Z. Liu, and Y. Zhang, “Adaptive Federated
Learning and Digital Twin for Industrial Internet of Things,” IEEE Pubudu N. Pathirana is a full Professor and the Director of Networked
Transactions on Industrial Informatics, pp. 1–1, Oct. 2020. Sensing and Control group at the School of Engineering, Deakin University,
[5] H. Liao, Z. Zhou, X. Zhao, L. Zhang, S. Mumtaz, A. Jolfaei, S. H. Geelong, Australia. He was a visiting professor at Yale University in 2009.
Ahmed, and A. K. Bashir, “Learning-Based Context-Aware Resource His current research interests include bio-medical assistive device design,
Allocation for Edge-Computing-Empowered Industrial IoT,” IEEE In- mobile/wireless networks, and Internet of Things.
ternet of Things Journal, vol. 7, no. 5, pp. 4260–4277, May 2020.
[6] H. Liao, Z. Zhou, X. Zhao, and Y. Wang, “Learning-Based Queue-
Aware Task Offloading and Resource Allocation for Space-Air-Ground- Aruna Seneviratne is currently a Foundation Professor of telecommunica-
Integrated Power IoT,” IEEE Internet of Things Journal, vol. 8, no. 7, tions with the University of New South Wales, Australia, where he holds
pp. 5250–5263, Apr. 2021. the Mahanakorn Chair of telecommunications. His current research interests
[7] J. Cao, K. Zhang, F. Wu, and S. Leng, “Learning Cooperation Schemes are in physical analytics: technologies that enable applications to interact
for Mobile Edge Computing Empowered Internet of Vehicles,” in Proc. intelligently and securely with their environment in real time.
IEEE Wireless Communications and Networking Conference (WCNC),
Seoul, Korea (South), May 2020, pp. 1–6.
[8] Y. Lu, X. Huang, Y. Dai, S. Maharjan, and Y. Zhang, “Differentially Jun Li received Ph. D degree in Electronic Engineering from Shanghai
Private Asynchronous Federated Learning for Mobile Edge Computing Jiao Tong University, China in 2009. His research interests include network
in Urban Informatics,” IEEE Transactions on Industrial Informatics, information theory, game theory, distributed intelligence, multiple agent
vol. 16, no. 3, pp. 2134–2143, Mar. 2020. reinforcement learning. He has co-authored more than 200 papers in IEEE
[9] Y. Song, T. Liu, T. Wei, X. Wang, Z. Tao, and M. Chen, “FDA3: journals and conferences, and holds 1 US patents and more than 10 Chinese
Federated Defense Against Adversarial Attacks for Cloud-Based IIoT patents in these areas. He was serving as an editor of IEEE Communication
Applications,” IEEE Transactions on Industrial Informatics, pp. 1–1, Letters and TPC member for several flagship IEEE conferences.
Jun. 2020.
[10] Y. Liu, Z. Ma, X. Liu, S. Ma, S. Nepal, and R. Deng, “Boosting
Privately: Privacy-Preserving Federated Extreme Boosting for Mobile Dusit Niyato (F’17) received the B.Eng. degree from the King Mongkuts
Crowdsensing,” arXiv:1907.10218, Apr. 2020. Institute of Technology Ladkrabang, Thailand, in 1999, and the Ph.D. de-
[11] Z. Du, C. Wu, T. Yoshinaga, K.-L. A. Yau, Y. Ji, and J. Li, “Federated gree from the University of Manitoba, Canada, in 2008. He is currently a
Learning for Vehicular Internet of Things: Recent Advances and Open Professor with the School of Computer Science and Engineering, Nanyang
Issues,” IEEE Open Journal of the Computer Society, vol. 1, pp. 45–61, Technological University, Singapore. His research interests are in the areas
May 2020. of energy harvesting for wireless communication, the Internet of Things, and
[12] A. Taik and S. Cherkaoui, “Electrical Load Forecasting Using Edge sensor networks.
Computing and Federated Learning,” in Proc. IEEE International Con-
ference on Communications (ICC), Dublin, Ireland, Jun. 2020, pp. 1–6.
[13] M. Hao, H. Li, G. Xu, Z. Liu, and Z. Chen, “Privacy-aware and H. Vincent Poor (F’87) is the Michael Henry Strater University Professor of
Resource-saving Collaborative Learning for Healthcare in Cloud Com- Electrical Engineering at Princeton University. His interests include informa-
puting,” in Proc. IEEE International Conference on Communications tion theory, machine learning and networks science, and their applications in
(ICC), Dublin, Ireland, Jun. 2020, pp. 1–6. wireless networks, energy systems, and related fields. Dr. Poor is a Member
[14] B. Liu, B. Yan, Y. Zhou, Y. Yang, and Y. Zhang, “Experiments of of the National Academy of Engineering and the National Academy of
Federated Learning for COVID-19 Chest X-ray Images,” Jul. 2020, Sciences, and a Foreign Member of the Chinese Academy of Sciences and
arXiv: 2007.05592. the Royal Society. He received the Marconi and Armstrong Awards of the
[15] Y. Jiang, H. Chen, M. H. Loew, and H. Ko, “COVID-19 CT Image IEEE Communications Society in 2007 and 2009, respectively, and the IEEE
Synthesis with a Conditional Generative Adversarial Network,” IEEE Alexander Graham Bell Medal in 2017.
Journal of Biomedical and Health Informatics, pp. 1–1, Dec. 2020.